Chapter 6. Mammograms. Neural Network Based Classification of. Majority of the work done in this area aims at detecting one or more of the three

Transcription

1 Chapter 6 Neural Network Based Classification of Mammograms A great deal of effort has been devoted to CAD in digital mammography to increase diagnostic accuracy as well as the reproducibility of mammographic interpretation. Majority of the work done in this area aims at detecting one or more of the three abnonnal structures in mammograms [10]: microcalcifications [32], [77], [90], [95], [99] circumscribed masses [57], [59], [69] and spiculated lesions [73], [76] which often characterize early breast cancer. Others have explored classifying breast lesions as benign or malignant [I10], [1 12]. There has also been work on distinguishing norrn/al regions from regions containing calcifications or masses. However, little work has been done on specifically characterizing nonnal mammograms [l20]. Treatment of cancer is most effective when it is detected early. The ability to reduce the proportion of 1nisses" in the cancer detection task enhances the chances of survival. Prescreening mammograms to identify the relatively large number of normal mammograms as well as areas of normal tissue in potentially abnormal mammograms will substantially reduce the workload of radiologists and increase the accuracy of their diagnosis in subtle cases.

2 123 Chapter 6. Neural Network based Classification of Mammograms In this chapter a neural network based classifier is developed for separating nonnal regions from potentially abnonnal regions in mammograms. Removing nomial background structures and normal linear markings from the mammograms under consideration, enhances the abnormal features. Statistical as well as structural features are extracted from these and the mammographic regions are classified into nonnal / abnormal classes with the help of a PNN classifier. 6.1 Normal Mammogram Characterization The problem of detecting normal mammograms is different from the detection of lesions and is not simply l - the detection of abnormal mammograms." Let P(cancer) be the probability of presence of cancer; P(normaI) be the probability of absence of cancer; P(image) be the probability of obtaining a specific mammogram; P(cancer/image) be the probability of cancer detection given a specific mammogram; P(image/cancer) be the probability of obtaining a specific mammogram given there is cancer and P(image/normal) be the probability of obtaining a specific mammogram given there is no cancer. In the context of a decision model using Bayes'.s rule in probability theory we have P(image / cancer)p(cancer) P(cancer / image) = _ (6.1) P(image) where P(image) = P(image/ cancer)p(cancer) + P(image/ normal )P(normal) (6.2) mcancer /image) = P(image/ cancer)p(cancer) (6.3) P(image / cancer)p(cancer) + P(image / normal )P(n0rmaI ) Detection is based on the posterior probability P (cancer/image), but the decision maybe highly sub-optimal if P (image/normal) is not known. For example, even if P (image/cancer) is small, the posterior probability can be large if P(image/normal) is close to zero. This means that misses can be avoided more easily, if

3 6. 2. Features of Normal Mammograms 129 the detection algorithms work synergistieally with algorithms for characterizing nonnal mammograms. Our approach to the nonnal mammogram recognition problem is based on normal tissue identification and removal, which is independent of the type of abnormalities that may exist in the mammogram. This approach also facilitates the classification of abnomialities, since suppressing normal background structures enhances the contrast and obviousness of abnormal structures. In addition, the r1-,~~.':='::1.= tissue characterization problem is fundamentally simpler and easier for computers to solve than the tumor detection problem, because the properties of images of nomial tissue are much simpler than the properties of images ofabnonnalities of various types, sizes, and stages of development. The classification is achieved by presenting the features from the residual image to a neural network classifier. 6.2 Features of Normal Mammograms Breast tissue composition varies with age and hormone levels in a woman. Generally, a young woman has denser or fibro-glandular breasts, which appear very white or "cloudy in a mammogram (Figure 3.2). Middle-aged women have a mixture of fibrous and glandular tissues (Figure 3.3). Their mammograms look black and white. In a mature breast, most of the fibrous tissue is replaced with fatty tissue. The mammograms tend to look black or gray (Figure 3.4). Completely normal mammograms may have entirely different appearances and hence a clear definition of normal mammograms is not easy [1], [2]. As there are no spikes corresponding to microcalcifications and no large bright areas corresponding to masses they have a lower overall density than abnormal ones. Normal regions have linear markings, which are shadows of ducts and connective tissue elements. These are distinct from spiculated or stellate lesions, in which linear markings radiate locally in all directions [120], [75](see fig 3.14). Normal linear markings in mammograms can be considered as straight-line segments of dimensions 1 to 2 mm or greater in length and

4 130 Chapter 6. Neural Network based Classification of Mammogram: 0.1 to lmm in width. Removal ofthe nomtal background and linear markings enhances the contrast and obviousncss of abnomtal structures making their detection easier. 6.3 Residual Image Generation Normal mammogram characterization is achieved by identifying and removing normal tissue structures. The mammograms obtained after removing the normal structures are called residual images. Accordingly, the residual image of a nonnal mammogram would be unifonnly dark and featureless whereas the residual image of an abnormal mammogram would show microcalcifrcations or masses against a featureless background. In this work two types of residual images are generated; one by removing the normal background structures and the other by removing the normal linear markings present in mammograms Removal of normal background Removal of normal background region helps in enhancing any abnormalities present in the ROI under consideration. Here, a WT based average subtraction technique is employed for background removal, as the WT is capable of separating small objects such as microcalcifications from large objects such as large background structures. It was found [214] that the resolution level 1 of the WT showed mainly the high frequency noise included in the mammogram, whereas levels 2 and 3 enhanced microcalcifications effectively. Levels greater than 3 showed a large correlation with the non-unifonn background. To enhance the abnormalities properly, the coarser background should be suppressed, without suppressing the finer abnormal features. This can be done using a smooth mother wavelet. As the wavelet becomes smoother, higher degree of background non-uniformity can be corrected but at the cost of localization property of

5 6.3. Residual Image Generation l3l the wavelet [26]. In this study. we have selected a mother wavelet. which has intermediate length and a high degree of smoothness, to extract subtle and small abnormalities and to suppress background structures effectively. Figures 6.1 to 6.7 show residual images generated by background-subtraction corresponding to a normal mammogram, and mammograms containing various types of abnonnalities like microcalcification. masses, asymmetry and architectural distortion. The images are decomposed to 4 levels using bior 6.5 wavelets and reconstructed after discarding the fourth level approximations. This helps to remove the normal background structures and projects abnormalities, if any. la) (b) Figure 6.1: Residual image generation (background removed) (21) Original normal mammogram b ) Residual image la) (b) Figure 6.2: Residual image generation (background removed) (a) Original image containing microcalcifications b) Residual image

6 132 Chapter 6. Neural Network based Class1_'ficaIi0n ()f.mamm0gram.s' (3) (b) Yigtirc 6.3: Residual image generation (background rcmo\'cd) (:1) (if)riginal image containing 21 circumscribed mass bl Residual image (at) (b) Figure 6.4: Residual ima ge 3eneration (background rcinmedl (fa; ()riginal image contain} ng a spiculatcd mass b) Residual image Detection and removal of linear markings Linear structure detection is a very basic, yet important problem in image processing and computer vision. It is also often the preprocessing step in other applications such as feature extraction, pattern recognition, and image enhancement. Since 1960s. the Hough transform [1]]. [215], [216] has been widely used for detecting lines in binary images. The Hough transform is fairly robust to noisy or missing data [216]. [217] and it can be

7 6. 3. Residual Image Generation 1L4.) '9.) (ill (b) Figure 6.5: Residual image generation (background rcnim ed) ta) Original image containing an asymmetr_v b) Residual imagc li- igure 6.6: Residual image generation (bacl\'grol1nd re-inmed) ta) Original image containing an architectural distortion b) Residual image easily extended to detect shapes other than straight lines, such as ellipse [218]. [219] and circular objects [220]. However, this requires increased storage and computational complexity [22]]. Another drawback of the Hough transfonn is that it is not suitable for direct use in grayscale images. It requires some preprocessing steps, such as edge detection and thresholding. to make an input grayscale image a binary pattern. In addition, the Hough transform does not provide the actual position of the line in the (x,y) plane [222]. but only the angle ofthe line Band its distance from the origin p. This is not sufficient for many applications.

8 134 Chapter 6. Neural Network based Classgfication Qf'Mamm0grams m) (b) figure 6.7: Resitlua ; image generation (background rc1no\'ctl> (Ll? %Ii rz T~_;inal iinage containing an ill delined mass b) Residual iinugc Lines are commonly viewed as extended or contiguous edges. Consequently. many line detection algorithms extract local edges first and then group them into more globally defined lines based on certain criteria [223], [224], [225], [226], [227], [228]. [229]. However. local edge operators usually enhance the noise and tend to generate dense edge maps due to their small spatial extent, which makes subsequent processing difficult [230]. Normal linear markings in digitized mammograms vary from 0.1 to 1mm in width. Hence, it is assumed that the basic characteristic of a line in a mammogram. regardless of its thickness, is that pixels on it have similar gray levels. To detect such lines. line detector should be capable of extracting lines with very different and irregular width, as well as curves. Hence the line detector proposed by Liu et al [i'i" l] is adopted in this work. which is described in Appendix A. Removal of nonnal linear markings reduces the possibility of wrong classification of nonnal mammograms into abnormal class. Residual images generated by removing normal linear markings using the above algorithm for different mammograms are shown in figures 6.8 to

9 6. 4 Neural Network Training and Testing Methodology l35 la) (b) Figure 6.8: Residual image generation (normal linear markings removed) (a) Original normal mammogram b) Residual image (a) (bl Figure 6.9: Residual image generation (normal linear markings removed) (a) Original image containing a microcalcification cluster b) Residual image 6.4 Neural Network Training and Testing Methodology Before a neural network model can be used as a pattern classifier, its structure has to be designed and trained. In this section the selection of training and testing data sets and training techniques are discussed.

10 136 Chapter 6. Neural Network based Classification 0fMamm0gram.s' (a) (b) Figure 6.10: Residual image generation (normal linear markings removed) (a) Original image containing a circumscribed mass b) Residual image (a) (b) Figure 6.1 1: Residual image generation (nonnal linear markings removed) (a) Original image containing a speculated lesion b) Residual image 6.4.] Training and Testing Data Sets The datasets used were the digitized mammographic data obtained from the freely available database provided by the MIAS [232]. The images in the database are digitized at 50-micron pixel edge, which are then reduced to 200-micron pixel edge and clipped or padded so that every image is having 1024 x 1024 pixels with the image

11 6.4 Neural Network Training and Testing Methodology I37 portion centered in the matrix. The origin of the coordinate system is the bottom-left comer. The accompanied Ground Truth contains details regarding the character of the background tissue, class and severity of the abnormality and x, y coordinate of its center and radii. The database contains 207 normal mammograms and IIS mammograms containing various abnonnalities such as calcification, circumscribed masses, spiculated masses, other ill-defined masses, architectural distortion and asymmetry. A detailed description ofthe data set is given in table 6.1. Normal 207 Type Total number Approximate range ofradius (in pixels) ofa circle enclosing the abnormality. Calcification 25 3 to 87 Circumscribed masses to l97 Spieulated lll-defined masses masses l I7 to to I23 I74 Total 322 Table 6.1: Training and testing data sets. Architectural distortion I9 23 to 117 AsymmetrL to l3l The mammograms in this database are arranged in pairs of films, where each pair represents the lefi (even filename numbers) and right mammograms (odd filename numbers) of a single patient. When calcifications are present, centre locations and radii apply to clusters rather than individual calcifications. In some cases calcifications are widely distributed throughout the image rather than concentrated at a single site. In these cases centre locations and radii are inappropriate and have been omitted Detection criteria The results of classification are expressed in tenns of three parameters, True Positive (TP), False Positive (FF) and False Negative (FN). A TP is obtained when a normal/abnormal R01 is correctly classified into normal/abnormal class. When a normal R01 is incorrectly classified as abnormal, it is defined as a FP. A FN is obtained when an abnormal R01 is incorrectly classified into normal class.

12 138 Chapter 6. Neural Network based Classification of Mammograms n-fold Cross Validation For the purpose of generalization, the cross-validation (CV) tcclmique [233], [234] is used. The cross validation method is a leave-one-out algorithm. For large data sets, this algorithm makes very heavy demands on computing resources. For instance, a Sun Ultra 1 workstation running a PNN simulator required three weeks continual processing to complete a cross validation exercise for the size/structure classification based on a database of around 5000 examples [233]. This is clearly impractical. A compromise solution is to use an n-fold cross validation. The database is divided, with random selection of examples, into n partitions (known sometimes as folds) of varying sizes. The literature on n-fold cross validatic: shows little discussion on the selection of n, with many researchers using values between 5 and 20 and the majority use n=l0 [233]. In this study, the partition sizes are varied from 25:75 training/validation to 50:50 training/validation with n=l Normal/abnormal classification based on statistical features Selection of Neural Network structure for classification The performance of four types of network architectures available in MATLAB, which are popular for classification applications, is compared here. The architectures selected for this study were the BPNN, the RBFN, the PNN and the Competitive network. The four statistical features described in section /4.1 l.1 viz. mean (,u), variance (02), skewness (#3) and kurtosis (#4 ) were fed to the input of the networks and the results are tabulated in table 6.2.

13 6. 5. Normal/abnormal classification based on statisticalfcamrcs 139 Type of No.of Original image Residual image with Residual image with network R015 background rcmovcd normal lines removed TP FP FN TP FP FN Tl FP FN BPNN Nonnal I Abnonnal Total I01 74 S9 234 RBFN Nonnal I65 Abnonnal Total PNN Nonnal Abnormal S Total Compet- Nonnal 2 I itive layer 165 Abnonnal I Total Table 6.2:Comparison of performance of different network architectures on nonnal/abnormal classification of mammographic data using statistical features. From the MIAS database 345 ROIs of size 256 x 256 were selected for this study. The selected ROIs include 221 normal ROIs and 124 R015 containing various abnonnalities. 111 R015 (55 abnormal ones and 56 normal ones) from the above are used for training the networks and the performance of the networks are tested using the remaining 234 ROIs (165 nonnal and 69 abnormal). The table clearly shows that best results are obtained for the PNN architecture. 1-lence, PNN is selected for the classification ofthe mammograms into normal and abnormal groups.

14 140 Chapter 6. Neural Network based Classification of Mammograms Feature selection A good feature must be sufficiently discriminating. However, in order to keep the classification problem tractable, the total number of features selected must be limited. Neural network technology offers techniques for selecting, developing, clustering and compressing features into a useful set. In this thesis, the problem of feature selection is attacked using a feature wrapper" approach. The guiding principle of this approach is that the features that can best be used for classification should be chosen. A consequence of this principle is that one must know exactly how the samples will be classified before feature selection can be done. A major advantage of the feature wrapper approach is accuracy, because the feature selection is tuned for the classification method. Another advantage is that the approach provides some protection against over fitting because of the internal cross validation employed by the jackknife approach. One drawback of feature wrapper method is that the method can be computationally intensive. The process of feature selection is described below [235]: (1) A candidate set of features is considered. a. The mammograms are divided into training and test sets. i. The classifier is trained on the training set of samples. ii. The classifier is used on the test set of samples. b. Step 1(a) is repeated with alternative divisions into training and test sets. c. The candidate feature set is evaluated using all classifications from 1(a) (i)-(ii). (2) Step 1 is repeated with another candidate feature set. The result of the feature wrapper approach for the statistical features,u, 02, #3 and,u4 derived from the original ROIS using PNN is tabulated in table 6.3. The PNN is trained and tested using the training and test data described in section The table shows that very low FP values are obtained for the feature kurtosis alone, but at the cost

15 6.5. Normal/abnormal classification based on statisricalfeaturcs l4l of a FN rate of 100%. This classification is intended as a pre-processing step for automated detection of breast cancer. Even though a slightly larger Fl rate can be tolerated, the F N rate should be as small as possible. Hence, best results are obtained for feature sets mean and variance, mean, skewness and kurtosis and all 4 together. But the classification result obtained here is far from the requirement. So the procedure is repeated on residual images obtained by subtracting the background and removing the nonnal linear markings. The results are tabulated in table 6.4 and 6.5. ll Selected 02 feature ll: #4 /11 1: 1: R. ~ ll» &&&&&&o CL}: 1: E R : /u /1; MM & & & &,u;& [11 /14 #4 /14 ;1; FP I02 2 I l I03 78 FN ! Table 6.3: Sensitivity for different features derived from the original image Selected feature it 0 #1 #4.u& xx in 0 0 #3 M A /1. 0" A 02 & & & & & 0 0 /1,,u; 0 xx; /14 ix; #4.114 & & & & p,& /11 #4 #4 /14 #4 FP l2l I ll FN I ll Table 6.4: Sensitivity for different features derived from the residual image (background removed) For residual images generated by background subtraction, lowest FN rate is produced by feature set mean, variance and skewness. But the FP rate of this feature set is unacceptably high. The feature sets mean, skewness and kurtosis and all 4 together provide acceptable Fl and FN rates. Since lesser number of features make the

16 142 Chapter 6. Neural Network ba.scd Classification of Mammograms computation easier, it is better to select the features mean, skewness and kurtosis for classification purpose, even though its FP rate is slightly higher compared to the result obtained by using all 4 features. The results obtained here are better than those obtained on features extracted from the original image. This is because of the enhancement occurred to the abnormalities due to the removal of normal background tissue from the images. Residual images formed by removing nonnal linear markings provide best results when the parameters mean and skewness are used (see table 6.5). Based on these observations three features mean, skewness and kurtosis are selected for classification of mammograms into normal and abnonnal classes. Selected feature it 0 I1: #4 t1 l1 #& 0 0 #1 I4 /4 I4 0 # & &,2, & & & o o p,&,u; o, 0} ll: /11 #4 /14 & &;t,,u, &;t,,u,& #1,u4 FP so as 79 so so FN Table 6.5: Sensitivity for different features derived from the residual image (normal lines removed) Classification results 3 sets of the feature vectors containing three elements each, the mean, skewness and kurtosis are extracted from each snippet of mammogram of size 256 x 256 and presented to 3 PNNS. The first set of feature vectors is extracted from the original mammograms, second from residual images obtained after the removal of normal background and third from residual images obtained after detection and removal of normal linear markings. The input layer of the PNN handles the features extracted from each ROI. As mentioned earlier, two output units denote the presence or absence of an abnormal tissue.

17 6.5. Normal/abnormal classi/zcation based on.r!ati.m'caifeaiurcs 143 The results for the three sets of features are tabulated in table % detection accuracy was obtained for the training set. Using the features from the original image a sensitivity of 52.7% for normal mammograms and 55% for abnonnal mammograms were obtained for the test data. A sensitivity of 57% for normal mammograms was obtained for the background-removed ease and 52% for residual images with linear markings removed. Both types of residual images produced a sensitivity of 75.4 % for the abnormal case. But none of these is sufficient for the first stage of an automated breast cancer detection system. Hence a two-step classification, as detailed below, was developed for increasing the sensitivity and specificity of detection. No.of Original image Residual image with Residual image ROls background removed with nonnal lines removed TP FP FN TP FP FN TP FP FN Nonnal I Abnonnal I7 Total I ! Table 6.6: Classification results for the 3 sets of feature vectors Step1: Classify the given mammogram into normal or abnormal groups using the features derived from the original mammograms. Step2: If the ROI is found to be normal, project any abnormality that can be present in it by removing the background and again classify. Else if it is abnormal, remove the normal linear markings that may be misunderstood as abnormalities and again classify. The result of this two-step classification is given in table 6.7. The detection sensitivity of the abnormal cases is 91% and that for the normal cases is only 58%. The

18 144 Chapter 6. Neural Network based Classification of Mammograms low sensitivity for the nonnal cases will not be much ofa problem as this is intended as a pre-processing stage only. A detailed classification result on the MIAS database is given in table 6.8. No.01 TP F P FN %Dctcction ROls Nomial I Abnonnal Total Table 6.7. Results of Statistical feature based classification ROls on Training set 56 S6 - - [00 Normal Test set Calcification 14 I I 00 circumscribed masses l2 l2 - - I00 Training Spiculated masses I I00 56! ill-defined masses l0 l0 - - I00 Architectural distortion Abnomlai Asymmetry Calcification I00 circumscribed masses 1 l 10 l 90.l Spiculated masses No.of TP FP FN %Detecti T St 53! ill-defined masses l 75 Architectural distortion 16 I Asymmetry l l Total Table 6.8. Detailed Results of Statistical feature based classification

19 6. 6. Normal/abnormal classification based on texmralfcarures Normal/abnormal classification based on textural features Selection of Neural Network structure for classification The perfomiancc of the same four types of networks as in section is estimated here to select the best network for classification based on textural features. Various texture features described in section were used for the evaluation and the!'f'sl'11 are tabulated in table 6.9. From the MIAS database, 332 R015 of size 256 x 'lC1LJ.('l1'1g 222 normal ones and 110 R015 containing various abnormalities iike microcalcifications, masses and architectural distortions are selected for this study. Type of No.01" Original image Residual image with Residual image network R015 background removed with nonnal lines removed TP FP FN TP Fl FN TP FP FN BPNN Normal Abnormal Total g _ Total Total RBFN Normal Abnormal PNN Normal Abnormal Competitive 168 Normal layer Abnormal Total Table 6.9: Comparison of performance of different network architectures on normal/abnormal classification ofmammographic data using textural features.

20 146 Chapter 6. Neural Network based Classification of Mammograms 109 (55 abnonnal ones and the rest normal) ROls from these were used to train the networks. Mammograms containing the abnormality, asymmetry is not included in this study Feature selection Ideally, the feature wrapper approach requires all possible candidate feature sets to be considered. But this is difficult even for modest number of candidate features. Hence, clever strategies are required to search through the space of feature sets. Xiong et al. evaluated two relatively simple search procedures, Sequential Forward Search (SFS) and Sequential Forward Floating Search (SFFS) [236]. The sequential forward search procedure is adopted here. Different steps involved in this are: (1) Choose the single best feature (2) Choose the best feature set of size two that includes the feature from (1) (3) Choose the best feature set of size three that includes the feature set from (2) (4) And so on. The textural features used for classification purpose are the Haralick s texture features described in section The results of various steps of SFS for feature selection on the test data set of 168 normal mammograms and 55 abnormal ones described in the previous section are tabulated below. Using this approach a feature set of 4 elements, which can best be used for classification is selected from the 10 features. The selected features are angular second moment or energy, entropy, correlation, and contrast. Table 6.10 shows that the best single feature for classification is correlation. The features that were not able to classify even the elements of the training set into their correct group were eliminated at this stage. The best result for a feature set size of 2 is obtained for the features correlation and entropy. Hence these two were selected as the best feature set of size 2 (see table 6.1 1). From table 6.12 the best feature set of size 3 is obtained by adding energy to that obtained in step 2. Following this procedure it can be

21 6. 6. Normal/abnormal cias.s'i'ficaiion based on Icxmralfcarures l47 seen that the best set of features of size 4 is correlation, entropy. energy and contrast (see table 6. l 3) and that for a 5 vector feature set includes local homogeneity in addition to the above 4. But when the 5-vector feature set is used, there is no improvement in performance over the 4-feature set. Hence, it is better to use the 4- feature set to reduce computational burden. energy entropy correlation contrast Sum of Local squares Homogenit variance y FP FN FP FN FP FN FP FN FP FN FP FN Original Back-ground I S subtracted Nonnal lines 165 I9 ll removed Table 6.10 Selection ofthe single best feature using SFS correlation correlation Correlation Correlation & Correlation & energy & entropy & contrast Local & Sum of Homogenity squares variance FP FN FP FN FP FN FP FN FP FN Original I Background subtracted Nomtal lines [39 ll removed Table Selection of the best feature set of size two using SFS correlation correlation correlation Correlation Tl entropy entropy & entropy & Local Entropy 8; Sum &energy contrast Homogenity of squares variance FP FN FP FN FP FN Fl FN Original Background lol subtracted Nonnal lines removed Table 6.12.Selection ofthe best feature set of size three using SFS

22 143 Chapter 6. Neural Network based Classification of Mammograms correlation entropy correlation entropy correlation entropy energy & energy & Local energy & Sum of contrast Homogenity squares variance Original FP FN I7 49 FP 68 F N 50 F P I9 F 55 N Background subtracted Nonnal lines I30 7 removed Table 6.13: Selection ofthe best feature set of size 4 using SFS F P FN FP FN correlation entropy energy correlation entropy energy contrast & Local Homogenity contrast & Sum of squares variance Original I7 49 I7 49 Background subtracted Nonnal lines removed Table 6. l4: Selection of the best feature set of size 5 using SFS Classification results Since the image matrix is discrete, the displacement vector used in the feature calculation was chosen to have the following phase and displacement values: (O,1), (45, 1), (9o,1), (135,1). The input layer of the PNN handles the four features extracted from each ROI. The three sets of features extracted from each ROI corresponding "'1 the original as well as the two residual images were tested using the PNN and the results are tabulated in table Though the false positive rate using the features from the original mammograms is less than 10%, the false negative rate is very high (94%), which cannot be tolerated. Best detection accuracy is got with the features obtained after removing normal lines. A recognition accuracy of 60% was obtained for the normal

23 6.6. Normal/abnormal ciassi_'/icarion based on Iexruralfcarurcs 149 mammograms and 94.2% for the abnomtal cases. For the entire data set an overall TP recognition score of 68.2% was obtained. The recognition score with features from the background-subtracted images is less. This is due to the fact that the texture infonnation is lost when the nomtal background is subtracted from the images. No of Original With background With nonnal lines mammo- subtracted removed grams TP FP FN TP FP FN TP F P FN Nonnal I68 l5l I l00 68 Abnonnal Total 223 I H I Table 6.15 Classification result using textural features Table 6.16 compares the perfonnance of the algorithm on the features obtained after removal of normal lines from the ROIs for different orientations. Best performance is attained for an orientation of 0. A detailed classification result is provided in table No of 0=o o=45 o=9o o=135 " mamm ogram TP rp FN rp FP FN TP FP FN T? F? FN Nonnal I68 I o Abnormal no Total I35 73 no L;.I Table 6.16 Classification result using textural features for different orientations

24 150 Chapter 6. Neural Network based Classification of Mammogram: No.of ROls TP FP FN / D l C!i " Training set Normal Test sct I68 I Calcification I00 circumscribed masses I00 Training Spiculatcd masses ! ill-defined masses I00 Architectural distortion J Abnormal Calcification 12 I2 - - I00 circumscribed masses I Spiculatcd masses TCSI 56! ill-defined masses Total Architectural ! distortion 68 l 3 I I Table 6.17:Detailed result of Classification for an orientation of Normal/abnormal classification based on both statistical and textural features The classification task is repeated using a combined feature set from the above sections. Residual images are formed by removing the normal linear markings. The 3 statistical features viz. mean, skewness and kurtosis and the four texture features viz. correlation entropy, energy and contrast described in previous two sections, are derived from the residual image and fed to the PNN for classification. 111 ROIs (55 abnormal ones and ones) from the MIAS database are used for training the PNN and its performance is tested using 236 ROIs (167 normal and 69 abnormal). The results are tabulated below in tables 6.18 and 6.19.

25 6. 8. Conclusion 151 No.of TI Fl FN %D l Cti0n Norma] ROls 167 lol Abnormal Total 236 I Table 6.18 Classification result using combined set of features. ROls on No.0! TP FP FN %Detecti Training set Nonnal Test set 167 I0 I Calcification 14 I4 - - I00 circumscribed masses 1 l l l - - I00 Training Spiculated masses l3 l ill-defined masses l0 I0 - - I00 Architectural distortion I00 Abnormal Asymmetry Calcification 20 I9 - I 95 E circumscribed masses Spiculated masses T65! 56! ill-defined masses Architectural distortion 16 I Asymmetry I Total "! Table 6.19 Detailed results of Classification using combined set of features 6.8 Conclusion PNN classifier for nomial/abnormal classification of digitized mammogram has been implemented based on statistical features, textural features and a combination of these.

26 152 Chapter 6. Neural Network based Classification of Mammograms A Tl identification rate of 58%, 59.5% and 60.5% were produced for the nonnal cases and 91.3%, 94.5% and 89.86% for the abnonnal cases respectively. The texture-based classifier was unable to classify mammograms having the abnonnality, asymmetry. The high TP rate of 94.5% for abnormal case using texture-based classifier is obtained excluding the mammograms containing asymmetry. As this is intended as the preprocessing stage of an automatic detection system, the major aim of this stage is to reduce the FN rate as far as possible. Hence, for further applications, the classifier based on statistical features is selected, even though it has slightly lower 'I";': E "'., :iv:':'.43 the other two.